3.3 · Action Chunks and Cross-Robot Normalization
Action chunks can improve temporal coherence and reduce inference frequency, but they introduce overlap and execution semantics. Normalization must preserve physical meaning across datasets and embodiments.
- Choose observation history and action horizon.
- Resolve overlapping action-chunk predictions.
- Normalize features and actions without losing units or compatibility.
Action chunks
A chunk policy predicts several future actions at once. This can coordinate contact and reduce expensive model calls. The system must define how a new chunk replaces, blends with, or is constrained by the remainder of the previous chunk.
History and horizon
Longer history may improve observability; longer horizons may improve coherence but increase prediction uncertainty. Select both with closed-loop latency and controller behavior in mind.
Normalization with provenance
Fit statistics on training data only and store them with the model. Standardization, bounded scaling, or robust quantile scaling must be tied to action units, controller limits, robot embodiment, and version.
Sequence length and normalization are physical interface choices, not merely training hyperparameters.
